8 x A100

AI & Data Science Cybersecurity
1

H Folks,

I have a demo with a Global Bank and I have qualified that they have gaps in their software security stack that we can fix with our peering exchange fingerprinting and Morpheus. I am looking to build a visual for them to help them get their heads around AI

I will demo the following:

  1. Real Time Asset Inventory
  2. Real Time Traffic Monitoring
  3. Virtual Packet Segmentation
  4. How all the above helps them close the gaps globally
  5. How building a digital twin will assist them to reduce risk and build a digital transition/AI-powered infrastructure programme.

I am hoping they will come on and work with us to then have a real customer use case with value add onto the Nvidia famework for GTC 2022.

I have the following sourced through our partners at Broadberry UK/US and would like to qualify if we have Triton support for it.

8x A100 AI all NVIDIA cards are interconnected via NVLINK total GPU memory is 640GB. 1TB RAM, CPU, NVMe’s + extra 8 PCIe x16 ports for our smartnic network connectivity (or direct-attached storage -

We noticed that the models you have available are for V100 on Sensitive Information Detection and Anomalous Behavior Profiling

I can see the

T4 GPU
$ ngc registry resource download-version
“ea-nvidia-morpheus/sid_bert_triton_package:sid-minibert-20210614-t4”

V100 GPU
$ ngc registry resource download-version
“ea-nvidia-morpheus/sid_bert_triton_package:sid-minibert-20210614-v100”

Is there a triton set for A100?

Any pointers greatly appreciated

2

Hi James,

The 0.1 EA release makes use of TensorRT 7.2 which does not support the A100, thus no pre-compiled engine was created at that time for use with Triton. The 0.2 EA release will use the 8.2 TensorRT libraries which will support A100. We may make use of JIT compilation of the ONNX model to TensorRT plan, as opposed to providing pre-compiled engines for each GPU model. Stay tuned.

\Pete

2 Likes
3

@jamestervit corrected some details in my reply.

1 Like
4

Hi Pete, thats really helpful thanks, I will go down the AWS route for my lab going into the trained models and start to build up the ;library on each sector.

As a side note:Is there any docker UI demo that I could spin up to support some pre-sales activity? I do have matplotlib visuals but is more playback as opposed to helping visualise the real-time, always a gamble between innovation and shrink wrap gaps.

thanks